Z-Image Turbo LoRA Training: Complete Guide with De-Distillation Adapter

Training custom LoRA models for Z-Image Turbo requires a specific approach due to its distilled architecture. This guide covers the de-distillation adapter technique developed by Ostris, training parameters, and deployment steps.

Quick Start: If you prefer a web-based solution, use our LoRA training tool to train custom Z-Image Turbo LoRAs directly in your browser—no local GPU required.

Understanding the De-Distillation Challenge

Z-Image Turbo is a step-distilled model—it achieves fast 8-step generation through a distillation process. Standard LoRA training breaks this distillation quickly, resulting in:

• Loss of speed benefits (no longer 8-step capable)
• Unpredictable quality degradation
• Artifacts in generated outputs

The de-distillation training adapter created by Ostris solves this problem.

How the Training Adapter Works

According to the Hugging Face engineering blog:

1. Adapter generation: Thousands of images were generated using Z-Image Turbo at various sizes and aspect ratios
2. Controlled distillation breakdown: A LoRA was trained on these images at low learning rate (1e-5), allowing distillation to break down in a controlled manner
3. Training on top: Your custom LoRA trains on top of this adapter, learning only your new content
4. Adapter removal: At inference time, remove the training adapter—your LoRA preserves distilled speed

This approach lets you train style, character, or concept LoRAs while maintaining 8-step Turbo generation.

Training Methods

Method 1: Web-Based Training (Recommended)

For those without local GPU resources, zimageturbo.com offers Z-Image Turbo LoRA training directly in your browser:

Benefits of our platform:

• No GPU required—training runs on cloud infrastructure
• Optimized parameters pre-configured for Z-Image Turbo
• Automatic de-distillation adapter integration
• Download trained LoRA weights directly

Method 2: Local Training (AI Toolkit)

For local training, use the Ostris AI Toolkit.

Hardware Requirements:

• Minimum: 12GB VRAM (RTX 3080, RTX 4070)
• Recommended: 16GB+ VRAM (RTX 4090, RTX 3090)
• CPU offloading available for lower VRAM systems

Installation:

git clone https://github.com/ostris/ai-toolkit
cd ai-toolkit
pip install -r requirements.txt

Download Training Adapter:

# Download from Hugging Face
huggingface-cli download ostris/zimage_turbo_training_adapterV2

Use V2 of the adapter for refined results.

Dataset Preparation

Quality training data is essential for good LoRA results.

Image Requirements

• Count: 5-15 images for characters, 15-25 for styles
• Resolution: 1024×1024 minimum (matches Z-Image Turbo's native resolution)
• Format: PNG or JPEG
• Variety: Include different angles, lighting, and contexts

Caption Files

Create a text file for each image with the same name:

training_images/
├── image001.png
├── image001.txt
├── image002.png
├── image002.txt

Caption format example:

a portrait of [trigger_word], detailed face, natural lighting

Use a consistent trigger word that will activate your LoRA during generation.

Folder Structure

my_lora_training/
├── config.yaml
├── images/
│   ├── 001.png
│   ├── 001.txt
│   ├── 002.png
│   └── 002.txt
└── output/

Training Parameters

Based on community-tested configurations from GitHub issues and tutorials:

Recommended Settings

Training Adapter Selection

Two options in AI Toolkit:

1. Z-Image Turbo W/ Training Adapter

   • Preserves 8-step Turbo speed
   • Best for shorter runs (< 5,000 steps)
   • Remove adapter at inference

2. Z-Image De-Turbo (De-Distilled)

   • No adapter needed at inference
   • Suitable for longer training runs
   • Slightly slower generation

For most use cases, option 1 (with training adapter) is recommended.

Running Training

Basic AI Toolkit Command

python run.py config.yaml

Sample config.yaml

job: extension
config:
  name: my_z_image_lora
  process:
    - type: sd_trainer
      training_folder: ./images
      base_model: Tongyi-MAI/Z-Image-Turbo
      training_adapter: ostris/zimage_turbo_training_adapterV2
      resolution: 1024
      train_batch_size: 1
      learning_rate: 1e-4
      max_train_steps: 3000
      network:
        type: lora
        rank: 16
        alpha: 16
      save_steps: 500
      output_dir: ./output

Adjust paths and parameters based on your dataset and hardware.

Training Time Estimates

Based on user reports:

Enable Low VRAM mode if you encounter memory errors on 12GB cards.

Using Your Trained LoRA

In ComfyUI

1. Copy LoRA file to ComfyUI/models/loras/
2. Add "Load LoRA" node to workflow
3. Connect to Z-Image Turbo model loader
4. Set LoRA strength (0.5-1.0)

In Python (Diffusers)

from diffusers import ZImagePipeline
import torch

pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.bfloat16
)
pipe.load_lora_weights("./my_lora.safetensors")
pipe.to("cuda")

# Generate with trigger word
image = pipe(
    prompt="a portrait of [trigger_word], detailed face",
    num_inference_steps=9,
    guidance_scale=0.0
).images[0]

Remember to include your trigger word in prompts.

Troubleshooting

LoRA Not Applying

• Verify file path is correct
• Check LoRA was trained on Z-Image Turbo (not FLUX or SD)
• Increase LoRA strength

Quality Degradation

• Training too long (distillation breakdown)
• Learning rate too high
• Reduce steps or use V2 training adapter

VRAM Errors

• Enable Low VRAM mode in AI Toolkit
• Reduce batch size to 1
• Use gradient checkpointing
• Consider web-based training on zimageturbo.com

Artifacts in Output

• Remove training adapter if using Method 1
• Check LoRA strength (reduce if artifacts appear)
• Verify training completed without errors

Best Practices

1. Start small: Test with 1,000 steps before committing to longer runs
2. Caption carefully: Good captions improve LoRA quality significantly
3. Use V2 adapter: The refined V2 training adapter produces better results
4. Monitor checkpoints: Save every 500 steps to find optimal training point
5. Test incrementally: Generate samples at each checkpoint to avoid overtraining

Related Resources

• Train your LoRA on zimageturbo.com — No GPU required
• Z-Image Turbo vs FLUX Comparison — Understand the technical differences
• ComfyUI Workflow Guide — Set up Z-Image Turbo locally
• Pricing Plans — View training credits and subscription options

Sources:

• Ostris Training Adapter - Hugging Face
• Training LoRA for Z-Image Turbo - Hugging Face Blog
• AI Toolkit - GitHub